Rodin - Gen 2.5 ExtremeHigh Text-to-3D

Rodin's Gen 2.5 line comes in three text-to-3D flavors - Fast, Regular, and ExtremeHigh - and this node is the luxury end of that range. The name is doing a lot of work: it tells you which tier of Hyper3D's cloud API gets called, and it's the top one. You type a prompt, it returns a textured 3D mesh, and none of the heavy lifting happens on your machine. This is a paid cloud service wearing a ComfyUI costume, which is both the appeal and the catch.

What it actually is

Rodin - Gen 2.5 ExtremeHigh Text-to-3D is a thin async wrapper around the Hyper3D Rodin API at api.hyper3d.com. Everything important - geometry generation, UVs, textures - happens on Hyper3D's servers. Your local GPU just supervises. That means it works on a potato laptop, but it also means you need an API key with credits, and every run is a job that takes real time to complete.

For context, ComfyUI's custom node scene is mostly about wrangling models you downloaded yourself, and the ecosystem's reputation for dependency hell is earned. This pack is the opposite: no model files at all. requirements.txt is literally aiohttp (plus asyncio, which is Python standard library anyway). You install it like any other node, but there's nothing to download.

How it works

Under the hood the node calls process_full_generation() from the pack's modules/ code: it submits a generation job with your prompt and settings, polls the API until the job finishes, then downloads the resulting files into a fresh timestamped folder inside ComfyUI/output/. The node's single output, model_path, is just a string pointing at that downloaded model - the .glb, .obj, or whatever format you picked. If you're following the example workflow, that path is where you'd hand things off to a viewer or a render pipeline.

The inputs that matter

Most of the defaults are fine, honestly. The ones you'll actually touch:

• api_key - wire this from the Rodin - API KEY node. Every generation node in this pack needs it; you can also paste the key straight in.
• prompt - your entire creative input. This is text-to-3D, so the prompt is the job.
• mesh_faces - default 1,000,000, range 20K to 2M. This is the quality dial; more faces means a denser mesh and a longer job.
• mesh_mode - Raw (raw triangles) or Quad (quad-dominant remesh, friendlier if you're going to sculpt or animate the result).
• Material_Type - PBR, Shaded, All, or None. Default is Shaded.
• geometry_file_format - glb (default), usdz, fbx, obj, or stl. glb is the safe default for almost everything.
• Creative - flips the geometry instruct mode to creative (a looser, more interpretive read of your prompt) instead of faithful.

There are also toggles for TAPose and is_micro, plus two optional inputs: bbox (feed it from the Rodin - BBox Controlnet node to force dimensions) and height_cm for physical scale. If you're just getting a first model out, ignore both.

Installing it

ComfyUI Manager is the easy path - search for ComfyUI-Rodin and install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Ravenmelt/ComfyUI-Rodin
# restart ComfyUI

It's a maintained fork of DeemosTech's original ComfyUI-Rodin. No model weights, no CUDA requirements - the only real dependency is aiohttp, which the pack installs from its requirements.txt on first load. Restart and you're done.

Where people get burned

None of this happens locally, so your failure modes are cloud-shaped. The README's error codes tell the story: MISS_API_KEY means the key didn't reach the node (double-check the Rodin - API KEY wiring), RODIN_ERROR is the API rejecting something, NO_JOBS_FOUND / MISS_SUBSCRIPTION_KEY mean the key or account is wrong, and a generic failure usually means no credits or no internet. Also note the README's warning: ComfyUI needs write permissions on the output directory for the download step.

The other thing people underestimate is latency. This is not a 10-second ComfyUI render - an ExtremeHigh job is a minutes-long cloud task, and the node blocks until it's done. Budget for it.

My take: ExtremeHigh is the tier you graduate to. If you're iterating on prompts, the Regular or Fast nodes are the ones you reach for; this one is for when a hero asset needs the best mesh Hyper3D can squeeze out, and you're OK paying for it.